Mean exact 50-move-rule utility
Experiment E029 / terminal paired aggregate (final combined result from matched positions)
Exact advice.
A bounded result.
A local causal effect of displaying two unranked exact-policy moves on exact 50-move-rule move utility for this cohort; no full-game transfer, Elo, learning, vision, search, or general chess-strength claim.
60 paired positions240/240 registered calls complete (all returned; no provider-missing calls)2 arms · 2 replicates per arm6 waves · max 10 concurrent chats
Terminal evidence / registered arm comparison
Exact top two mean utility
Both arms use the same unrestricted move-string tool. The registered difference is whether two unranked exact-policy moves are shown in the information panel.
Exact top two / mean utility0.495833Exact 50-move-rule utility
Paired utility effect+7.0833 ppExact top two minus board only
Whole-position 95% interval+2.9167 to +11.6667 ppBootstrap uncertainty
Paired sign-flipp=0.0032399676Two-sided registered statistic
Protocol scopeidentical unrestricted move-string tool in both arms; any legal move allowed.
Terminal aggregate / qualification attached
+7.0833 pp
Across 60 complete paired positions, mean exact 50-move-rule utility (a 0-to-1 score from the exact rule's outcome labels) was 0.425000 board only and 0.495833 with exact top-two advice. The whole-position 95% bootstrap interval (a resampling-based uncertainty range over positions) was +2.9167 pp to +11.6667 pp, with paired sign-flip p-value (a matched-position test that flips each difference's sign)=0.0032399676.
Analysis qualification (rules set before collection; analyzer caveat retained). confirmatory statistics implement prospectively registered rules, but the executable analyzer was completed after terminal data and was not hash-bound in the execution lock
Analysis boundary. The registered statistics and decision rule classify this cohort as positive, but the executable analyzer was completed after terminal data and was not hash-bound before collection. Analyzer pre-hash lock means the executable analysis code was fixed before collection; this record discloses that it was not.
Mean utility with two moves shown
33.3333% to 96.6667%
240/240 calls · 0 provider-missing
Failure-preserving chronology (the events in order)
Seven protocol gates.
Each failure retained.
- 01
Original source version failed closed
NO-GO: one preregistered material-by-color outcome stratum was infeasible; zero Luna calls.
- 02
Source Selection Amendment 1
Post-tablebase-label and pre-model-outcome min-cost-flow selection preserved the feasible registered margins without generating or querying new positions.
- 03
Amended corpus qualified
The 60-position corpus has complete all-legal labels and reproduced exactly on a second provider-free parse, with a disclosed single-source limitation.
- 04
Execution interface incident
Returned action strings reached a UCI-only event boundary even though the prompt exposed UCI and SAN through an unrestricted string tool; no scored event preceded the stop.
- 05
Recovery Amendment 1 rejected
A request-body replay guard incorrectly blocked legitimate byte-identical future replicates; preflight rejected it with zero calls under the draft.
- 06
Recovery Amendment 2 accepted
Attempt identities protect against replay; legal UCI is canonicalized, otherwise only unambiguous legal SAN is converted, uniformly across both arms.
- 07
Terminal result admitted separately
A versioned content-addressed terminal aggregate is public through a separate reviewed sanitizer, with the post-run analyzer implementation-lock deviation disclosed.
Protocol scope and uncertainty
One defined cohort.
Exact-advice effect on 50-move-rule move utility for one Luna snapshot and one frozen synthetic five-piece population.
Model. openai/gpt-5.6-luna / openai/gpt-5.6-luna-20260709.
Scoring. complete frozen all-legal five-category Syzygy labels mapped to exact 50-move-rule utility.
Uncertainty. The whole-position 95% interval was +2.9167 pp to +11.6667 pp; paired sign-flip p=0.0032399676; paired median utility effect 0.000000; unresolved utility bounds (the range after allowing registered missing-cell extremes) 0.070833 to 0.070833.
Failure boundaries
What this does not establish.
- full-game transfer or Elo improvement
- unassisted Luna chess strength
- base-model learning or general agentic improvement
- human-like reasoning, better vision, or better search
- benefit outside five-piece tablebase positions or for another model
Public evidence boundary
Publish the protocol.
Withhold the positions.
Included
aggregate design counts
amendment chronology
claim limits
content identities
Excluded
positions
moves
prompts
responses
request identifiers
provider metadata
live progress
tokens
costs
account state
interim outcomes
Read the terminal result, then the failure-preserving account of how the protocol reached a valid cohort without quietly erasing its amendments.